Benchmark RadarAI BENCHMARK PROFILE
REKEY
ReKey is a live benchmark protocol that regenerates visual keys in VQA images at evaluation time, creating fresh instances with new answers, to combat data leakage and memorization.
- Released
- 2026-06-17
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Static benchmarks become contaminated over time; ReKey provides a contamination-resilient evaluation framework, ensuring scores reflect genuine visual ability.
Motivation
Static visual question answering (VQA) benchmarks age quickly: Once the items leak into training corpora, scores can reflect memorization rather than genuine visual ability, thus obscuring real progress.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.